arch/s390/mm/vmem.c

   1 /*
   2  *    Copyright IBM Corp. 2006
   3  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
   4  */
   5
   6 #include <linux/bootmem.h>
   7 #include <linux/pfn.h>
   8 #include <linux/mm.h>
   9 #include <linux/module.h>
  10 #include <linux/list.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/slab.h>
  13 #include <asm/pgalloc.h>
  14 #include <asm/pgtable.h>
  15 #include <asm/setup.h>
  16 #include <asm/tlbflush.h>
  17 #include <asm/sections.h>
  18
  19 static DEFINE_MUTEX(vmem_mutex);
  20
  21 struct memory_segment {
  22         struct list_head list;
  23         unsigned long start;
  24         unsigned long size;
  25 };
  26
  27 static LIST_HEAD(mem_segs);
  28
  29 static void __ref *vmem_alloc_pages(unsigned int order)
  30 {
  31         if (slab_is_available())
  32                 return (void *)__get_free_pages(GFP_KERNEL, order);
  33         return alloc_bootmem_pages((1 << order) * PAGE_SIZE);
  34 }
  35
  36 static inline pud_t *vmem_pud_alloc(void)
  37 {
  38         pud_t *pud = NULL;
  39
  40 #ifdef CONFIG_64BIT
  41         pud = vmem_alloc_pages(2);
  42         if (!pud)
  43                 return NULL;
  44         clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
  45 #endif
  46         return pud;
  47 }
  48
  49 static inline pmd_t *vmem_pmd_alloc(void)
  50 {
  51         pmd_t *pmd = NULL;
  52
  53 #ifdef CONFIG_64BIT
  54         pmd = vmem_alloc_pages(2);
  55         if (!pmd)
  56                 return NULL;
  57         clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
  58 #endif
  59         return pmd;
  60 }
  61
  62 static pte_t __ref *vmem_pte_alloc(unsigned long address)
  63 {
  64         pte_t *pte;
  65
  66         if (slab_is_available())
  67                 pte = (pte_t *) page_table_alloc(&init_mm, address);
  68         else
  69                 pte = alloc_bootmem(PTRS_PER_PTE * sizeof(pte_t));
  70         if (!pte)
  71                 return NULL;
  72         clear_table((unsigned long *) pte, _PAGE_TYPE_EMPTY,
  73                     PTRS_PER_PTE * sizeof(pte_t));
  74         return pte;
  75 }
  76
  77 /*
  78  * Add a physical memory range to the 1:1 mapping.
  79  */
  80 static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
  81 {
  82         unsigned long end = start + size;
  83         unsigned long address = start;
  84         pgd_t *pg_dir;
  85         pud_t *pu_dir;
  86         pmd_t *pm_dir;
  87         pte_t *pt_dir;
  88         int ret = -ENOMEM;
  89
  90         while (address < end) {
  91                 pg_dir = pgd_offset_k(address);
  92                 if (pgd_none(*pg_dir)) {
  93                         pu_dir = vmem_pud_alloc();
  94                         if (!pu_dir)
  95                                 goto out;
  96                         pgd_populate(&init_mm, pg_dir, pu_dir);
  97                 }
  98                 pu_dir = pud_offset(pg_dir, address);
  99 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
 100                 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
 101                     !(address & ~PUD_MASK) && (address + PUD_SIZE <= end)) {
 102                         pud_val(*pu_dir) = __pa(address) |
 103                                 _REGION_ENTRY_TYPE_R3 | _REGION3_ENTRY_LARGE |
 104                                 (ro ? _REGION_ENTRY_RO : 0);
 105                         address += PUD_SIZE;
 106                         continue;
 107                 }
 108 #endif
 109                 if (pud_none(*pu_dir)) {
 110                         pm_dir = vmem_pmd_alloc();
 111                         if (!pm_dir)
 112                                 goto out;
 113                         pud_populate(&init_mm, pu_dir, pm_dir);
 114                 }
 115                 pm_dir = pmd_offset(pu_dir, address);
 116 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
 117                 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
 118                     !(address & ~PMD_MASK) && (address + PMD_SIZE <= end)) {
 119                         pmd_val(*pm_dir) = __pa(address) |
 120                                 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
 121                                 (ro ? _SEGMENT_ENTRY_RO : 0);
 122                         address += PMD_SIZE;
 123                         continue;
 124                 }
 125 #endif
 126                 if (pmd_none(*pm_dir)) {
 127                         pt_dir = vmem_pte_alloc(address);
 128                         if (!pt_dir)
 129                                 goto out;
 130                         pmd_populate(&init_mm, pm_dir, pt_dir);
 131                 }
 132
 133                 pt_dir = pte_offset_kernel(pm_dir, address);
 134                 pte_val(*pt_dir) = __pa(address) | (ro ? _PAGE_RO : 0);
 135                 address += PAGE_SIZE;
 136         }
 137         ret = 0;
 138 out:
 139         flush_tlb_kernel_range(start, end);
 140         return ret;
 141 }
 142
 143 /*
 144  * Remove a physical memory range from the 1:1 mapping.
 145  * Currently only invalidates page table entries.
 146  */
 147 static void vmem_remove_range(unsigned long start, unsigned long size)
 148 {
 149         unsigned long end = start + size;
 150         unsigned long address = start;
 151         pgd_t *pg_dir;
 152         pud_t *pu_dir;
 153         pmd_t *pm_dir;
 154         pte_t *pt_dir;
 155         pte_t  pte;
 156
 157         pte_val(pte) = _PAGE_TYPE_EMPTY;
 158         while (address < end) {
 159                 pg_dir = pgd_offset_k(address);
 160                 if (pgd_none(*pg_dir)) {
 161                         address += PGDIR_SIZE;
 162                         continue;
 163                 }
 164                 pu_dir = pud_offset(pg_dir, address);
 165                 if (pud_none(*pu_dir)) {
 166                         address += PUD_SIZE;
 167                         continue;
 168                 }
 169                 if (pud_large(*pu_dir)) {
 170                         pud_clear(pu_dir);
 171                         address += PUD_SIZE;
 172                         continue;
 173                 }
 174                 pm_dir = pmd_offset(pu_dir, address);
 175                 if (pmd_none(*pm_dir)) {
 176                         address += PMD_SIZE;
 177                         continue;
 178                 }
 179                 if (pmd_large(*pm_dir)) {
 180                         pmd_clear(pm_dir);
 181                         address += PMD_SIZE;
 182                         continue;
 183                 }
 184                 pt_dir = pte_offset_kernel(pm_dir, address);
 185                 *pt_dir = pte;
 186                 address += PAGE_SIZE;
 187         }
 188         flush_tlb_kernel_range(start, end);
 189 }
 190
 191 /*
 192  * Add a backed mem_map array to the virtual mem_map array.
 193  */
 194 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
 195 {
 196         unsigned long address, start_addr, end_addr;
 197         pgd_t *pg_dir;
 198         pud_t *pu_dir;
 199         pmd_t *pm_dir;
 200         pte_t *pt_dir;
 201         int ret = -ENOMEM;
 202
 203         start_addr = (unsigned long) start;
 204         end_addr = (unsigned long) (start + nr);
 205
 206         for (address = start_addr; address < end_addr;) {
 207                 pg_dir = pgd_offset_k(address);
 208                 if (pgd_none(*pg_dir)) {
 209                         pu_dir = vmem_pud_alloc();
 210                         if (!pu_dir)
 211                                 goto out;
 212                         pgd_populate(&init_mm, pg_dir, pu_dir);
 213                 }
 214
 215                 pu_dir = pud_offset(pg_dir, address);
 216                 if (pud_none(*pu_dir)) {
 217                         pm_dir = vmem_pmd_alloc();
 218                         if (!pm_dir)
 219                                 goto out;
 220                         pud_populate(&init_mm, pu_dir, pm_dir);
 221                 }
 222
 223                 pm_dir = pmd_offset(pu_dir, address);
 224                 if (pmd_none(*pm_dir)) {
 225 #ifdef CONFIG_64BIT
 226                         /* Use 1MB frames for vmemmap if available. We always
 227                          * use large frames even if they are only partially
 228                          * used.
 229                          * Otherwise we would have also page tables since
 230                          * vmemmap_populate gets called for each section
 231                          * separately. */
 232                         if (MACHINE_HAS_EDAT1) {
 233                                 void *new_page;
 234
 235                                 new_page = vmemmap_alloc_block(PMD_SIZE, node);
 236                                 if (!new_page)
 237                                         goto out;
 238                                 pmd_val(*pm_dir) = __pa(new_page) |
 239                                         _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE;
 240                                 address = (address + PMD_SIZE) & PMD_MASK;
 241                                 continue;
 242                         }
 243 #endif
 244                         pt_dir = vmem_pte_alloc(address);
 245                         if (!pt_dir)
 246                                 goto out;
 247                         pmd_populate(&init_mm, pm_dir, pt_dir);
 248                 } else if (pmd_large(*pm_dir)) {
 249                         address = (address + PMD_SIZE) & PMD_MASK;
 250                         continue;
 251                 }
 252
 253                 pt_dir = pte_offset_kernel(pm_dir, address);
 254                 if (pte_none(*pt_dir)) {
 255                         unsigned long new_page;
 256
 257                         new_page =__pa(vmem_alloc_pages(0));
 258                         if (!new_page)
 259                                 goto out;
 260                         pte_val(*pt_dir) = __pa(new_page);
 261                 }
 262                 address += PAGE_SIZE;
 263         }
 264         memset(start, 0, nr * sizeof(struct page));
 265         ret = 0;
 266 out:
 267         flush_tlb_kernel_range(start_addr, end_addr);
 268         return ret;
 269 }
 270
 271 /*
 272  * Add memory segment to the segment list if it doesn't overlap with
 273  * an already present segment.
 274  */
 275 static int insert_memory_segment(struct memory_segment *seg)
 276 {
 277         struct memory_segment *tmp;
 278
 279         if (seg->start + seg->size > VMEM_MAX_PHYS ||
 280             seg->start + seg->size < seg->start)
 281                 return -ERANGE;
 282
 283         list_for_each_entry(tmp, &mem_segs, list) {
 284                 if (seg->start >= tmp->start + tmp->size)
 285                         continue;
 286                 if (seg->start + seg->size <= tmp->start)
 287                         continue;
 288                 return -ENOSPC;
 289         }
 290         list_add(&seg->list, &mem_segs);
 291         return 0;
 292 }
 293
 294 /*
 295  * Remove memory segment from the segment list.
 296  */
 297 static void remove_memory_segment(struct memory_segment *seg)
 298 {
 299         list_del(&seg->list);
 300 }
 301
 302 static void __remove_shared_memory(struct memory_segment *seg)
 303 {
 304         remove_memory_segment(seg);
 305         vmem_remove_range(seg->start, seg->size);
 306 }
 307
 308 int vmem_remove_mapping(unsigned long start, unsigned long size)
 309 {
 310         struct memory_segment *seg;
 311         int ret;
 312
 313         mutex_lock(&vmem_mutex);
 314
 315         ret = -ENOENT;
 316         list_for_each_entry(seg, &mem_segs, list) {
 317                 if (seg->start == start && seg->size == size)
 318                         break;
 319         }
 320
 321         if (seg->start != start || seg->size != size)
 322                 goto out;
 323
 324         ret = 0;
 325         __remove_shared_memory(seg);
 326         kfree(seg);
 327 out:
 328         mutex_unlock(&vmem_mutex);
 329         return ret;
 330 }
 331
 332 int vmem_add_mapping(unsigned long start, unsigned long size)
 333 {
 334         struct memory_segment *seg;
 335         int ret;
 336
 337         mutex_lock(&vmem_mutex);
 338         ret = -ENOMEM;
 339         seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 340         if (!seg)
 341                 goto out;
 342         seg->start = start;
 343         seg->size = size;
 344
 345         ret = insert_memory_segment(seg);
 346         if (ret)
 347                 goto out_free;
 348
 349         ret = vmem_add_mem(start, size, 0);
 350         if (ret)
 351                 goto out_remove;
 352         goto out;
 353
 354 out_remove:
 355         __remove_shared_memory(seg);
 356 out_free:
 357         kfree(seg);
 358 out:
 359         mutex_unlock(&vmem_mutex);
 360         return ret;
 361 }
 362
 363 /*
 364  * map whole physical memory to virtual memory (identity mapping)
 365  * we reserve enough space in the vmalloc area for vmemmap to hotplug
 366  * additional memory segments.
 367  */
 368 void __init vmem_map_init(void)
 369 {
 370         unsigned long ro_start, ro_end;
 371         unsigned long start, end;
 372         int i;
 373
 374         ro_start = PFN_ALIGN((unsigned long)&_stext);
 375         ro_end = (unsigned long)&_eshared & PAGE_MASK;
 376         for (i = 0; i < MEMORY_CHUNKS && memory_chunk[i].size > 0; i++) {
 377                 if (memory_chunk[i].type == CHUNK_CRASHK ||
 378                     memory_chunk[i].type == CHUNK_OLDMEM)
 379                         continue;
 380                 start = memory_chunk[i].addr;
 381                 end = memory_chunk[i].addr + memory_chunk[i].size;
 382                 if (start >= ro_end || end <= ro_start)
 383                         vmem_add_mem(start, end - start, 0);
 384                 else if (start >= ro_start && end <= ro_end)
 385                         vmem_add_mem(start, end - start, 1);
 386                 else if (start >= ro_start) {
 387                         vmem_add_mem(start, ro_end - start, 1);
 388                         vmem_add_mem(ro_end, end - ro_end, 0);
 389                 } else if (end < ro_end) {
 390                         vmem_add_mem(start, ro_start - start, 0);
 391                         vmem_add_mem(ro_start, end - ro_start, 1);
 392                 } else {
 393                         vmem_add_mem(start, ro_start - start, 0);
 394                         vmem_add_mem(ro_start, ro_end - ro_start, 1);
 395                         vmem_add_mem(ro_end, end - ro_end, 0);
 396                 }
 397         }
 398 }
 399
 400 /*
 401  * Convert memory chunk array to a memory segment list so there is a single
 402  * list that contains both r/w memory and shared memory segments.
 403  */
 404 static int __init vmem_convert_memory_chunk(void)
 405 {
 406         struct memory_segment *seg;
 407         int i;
 408
 409         mutex_lock(&vmem_mutex);
 410         for (i = 0; i < MEMORY_CHUNKS; i++) {
 411                 if (!memory_chunk[i].size)
 412                         continue;
 413                 if (memory_chunk[i].type == CHUNK_CRASHK ||
 414                     memory_chunk[i].type == CHUNK_OLDMEM)
 415                         continue;
 416                 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 417                 if (!seg)
 418                         panic("Out of memory...\n");
 419                 seg->start = memory_chunk[i].addr;
 420                 seg->size = memory_chunk[i].size;
 421                 insert_memory_segment(seg);
 422         }
 423         mutex_unlock(&vmem_mutex);
 424         return 0;
 425 }
 426
 427 core_initcall(vmem_convert_memory_chunk);